I Can't Stay Quiet on Broadcom (AVGO) vs Nvidia Stock (NVDA) Any Longer

Key Concepts

Broadcom: A technology company specializing in semiconductor and infrastructure software solutions.
Nvidia: A leading company in GPUs (Graphics Processing Units) and AI hardware.
Custom ASICs (Application-Specific Integrated Circuits): Chips designed for a specific purpose, offering high performance per dollar and per watt for targeted workloads.
GPUs (Graphics Processing Units): Versatile processors capable of handling a wide range of computational tasks, particularly effective for training large AI models and complex graphics rendering.
Ethernet Switching Chips: Components that manage data flow within data centers, crucial for networking and communication.
Infiniband: A high-performance interconnect technology often used in AI data centers for high-speed communication between GPUs.
VMware: A company providing virtualization software, enabling multiple operating systems and applications to run on the same physical hardware.
Semantic: A company offering cybersecurity software solutions.
Titan XPU: Broadcom's custom processing unit designed for OpenAI, optimized for inference workloads.
Inference Workloads: The process of using a trained AI model to make predictions or generate outputs.
Training Workloads: The process of teaching an AI model by feeding it large amounts of data.
Data Centers: Facilities that house computer systems and associated components, such as telecommunications and storage systems.
Hyperscalers: Large cloud computing providers like Google, Amazon, and Microsoft.
Operating Income: A company's profit after deducting operating expenses from its gross profit.
Gross Profit: Revenue minus the cost of goods sold.
Recurring Revenue: Revenue that a company can expect to receive consistently over time.

Broadcom's Business and Competitive Landscape

Broadcom operates with two primary business units: Semiconductor Solutions (57% of revenue) and Infrastructure Software (43% of revenue). The company has a history of strategic acquisitions, notably VMware and Symantec, to build an end-to-end hardware and software portfolio for data centers.

Semiconductor Solutions

This segment focuses on designing custom AI accelerators, networking chips, and other networking devices. Broadcom holds a dominant position in the data center Ethernet switching market, with approximately 90% market share. Their Tomahawk chips are widely used by data center switch manufacturers like Arista Networks and Juniper, and even integrated into some Cisco switches. These chips are designed with AI and machine learning in mind, supporting features like:

Tightly coupled GPU clusters: Enabling efficient communication between multiple GPUs.
Ultra-low latency: Minimizing delays in data transmission.
Distributed training workloads: Facilitating the training of AI models across thousands of accelerators.
Co-packaged optics: Integrating optical components directly into the chip package to reduce signal loss and power consumption.

The widespread adoption of Ethernet in data centers (over 90% of global data center infrastructure) positions Broadcom favorably, especially as AI workloads running on Ethernet are steadily growing (currently 20-30%). The company's established install base and economies of scale make it challenging for new competitors to offer comparable performance or cost advantages.

Infrastructure Software

This unit largely comprises revenue from acquired companies:

VMware: Provides virtualization software that acts as an "operating system" for cloud and corporate data centers. Its flagship products, vSphere and VMware Cloud Foundation, enable companies to manage virtual machines, split resources (servers, storage, networking), and move applications between on-premises and cloud environments. The acquisition of VMware in 2023 has provided Broadcom with access to thousands of enterprise customers, stable recurring subscription revenue with high margins, and control over critical IT infrastructure.
Symantec: Offers cybersecurity solutions to protect businesses from threats like malware, ransomware, and data breaches. Their products cover endpoint protection (laptops, desktops, servers, smartphones), data loss prevention, and zero-trust gateways.

Partnership with OpenAI

A significant development is Broadcom's partnership with OpenAI to design the "Titan XPU," a custom processing unit optimized for inference workloads, particularly for models like ChatGPT. Mass production is slated to begin in 2026. These chips are exclusively for OpenAI's internal use, similar to custom silicon developed by Google (TPUs) and Meta (MTIA).

Key aspects of the Titan XPU:

Optimization for Inference: Designed to handle specific tasks like text-to-text, speech-to-text, text-to-speech, and code generation, which constitute the majority of OpenAI's inference needs.
Performance Advantages: Expected to offer 2-5 times better performance per dollar and per watt compared to Nvidia's GPUs for these specific workloads.
Focus on Text-Based Workloads: Tailored for processing text, a fundamental aspect of many AI applications.

This strategy aligns with the trend of major tech companies developing custom ASICs to complement, rather than replace, Nvidia's GPUs. The goal is to offload specific, high-demand inference tasks to these specialized chips, allowing Nvidia's GPUs to focus on the most computationally intensive and complex workloads, such as training massive AI models, rendering graphics, and running simulations.

Financial Performance and Outlook

Broadcom reported $16 billion in revenue for the latest quarter, a 22% year-over-year increase. Revenue specifically from AI grew by an impressive 63% year-over-year to $5.2 billion, with expectations of $6.2 billion in AI revenue for the next quarter. This rapid revenue ramp is comparable to Nvidia's growth trajectory.

However, the company faces challenges regarding its overall margins. Despite revenue and gross profit increases, operating income remained flat due to:

Higher Unit Costs and Lower Pricing for Custom Chips: New custom chips often involve new supply chains and are sold in bulk to hyperscalers at discounted prices, leading to higher production costs and lower per-chip revenue.
Increased Operating Expenses: A 9% year-over-year increase in operating expenses is attributed to significant investments in R&D, engineering, and sales to support new AI products.
VMware Integration Costs: The recent acquisition of VMware incurs restructuring and integration costs.

Competition with Nvidia and Market Dynamics

The video argues that Broadcom is not directly competing with Nvidia to replace its GPUs. Instead, Broadcom's custom ASICs are designed to address specific, high-demand inference workloads, particularly text-based ones. This approach aims to alleviate supply constraints on Nvidia's GPUs, which have been consistently sold out.

Nvidia, in response, is also evolving its offerings. The announcement of the "Rubin CPX" chip, designed for massive context inference, demonstrates Nvidia's strategy to enhance its GPU capabilities for specific AI tasks while also supporting both Infiniband and Spectrum X Ethernet.

The overall AI market is projected to grow significantly, with industry analysts expecting it to nearly 20x in size over the next nine years, representing a compound annual growth rate (CAGR) of close to 40% through 2034. This substantial growth suggests ample room for both Nvidia's GPUs and Broadcom's custom ASICs to thrive.

Investment Perspective

The speaker believes Nvidia will remain the "king of the entire AI revolution" due to the increasing complexity of AI models, the blurring lines between training and inference, and the need for flexible chips like GPUs. Nvidia's comprehensive hardware ecosystem, CUDA platform, and software services are considered significant competitive advantages.

However, the speaker acknowledges the disruptive potential of fundamentally different technologies, such as ASICs, for specific, high-demand workloads. Given the rapid growth of the AI market and the complementary nature of GPUs and ASICs, the speaker is moving Broadcom stock up three spots on their "get rich without getting lucky in 2025" list, placing it just below Nvidia. This strategy aims to capture growth from both leading companies, anticipating that both will succeed in the expanding AI landscape.

Notable Quotes

"99% of all internet traffic touches at least one Broadcom chip."
"The only reason that Nvidia's Infiniband doesn't have a 100% market share in AI data centers today is because Broadcom is so good at supporting distributed AI workloads over Ethernet."
"The goal of all of these custom chips isn't to disrupt NVIDIA at all. It's to make sure that Nvidia's GPUs are being used for the most computensive workloads and the most costefficient tasks."
"Today, the AI market is growing insanely fast, which means there's more than enough room for both GPUs and AS6, and I want to capture as much of that growth as I can."
"The best investment you can make is in you." (Attributed to Alex, Tickerol U)